L2QA: Long Legal Article Question Answering with Cascaded Key Segment Learning

Abstract

Evidences in Legal Question Answering (LQA) help infer accurate answers. Current sentence-level evidence extraction based methods may lose the discourse coherence of legal articles since they tend to make the extracted sentences scattered over an article. To this end, this paper proposes a cascaded key segment learning enhanced framework for $${\textbf {L}}$$ ong $${\textbf {L}}$$ egal article $${\textbf {Q}}$$ uestion $${\textbf {A}}$$ nswering, namely $${\textbf {L2QA}}$$ . The framework consists of three cascaded modules: Sifter, Reader, and Responder, which first transfers a long legal article into segments and each segment is inherent in the discourse coherence from consecutive sentences. And then, the Sifter is trained by automatically sifting out key segments in an iterative answer-guided coarse-to-fine way. The Reader utilizes a range of co-attention and self-attention mechanisms to obtain the semantic representations of the question and key segments. Finally, the Responder predicts final answers in a cascaded manner, identifying where the answer is located. Conducted on CAIL 2021 Law MRC dataset, our L2QA achieves 83.1 Macro-F1 and 65.8 EM and outperforms a state-of-the-art legal QA model by 4.1% and 9.1%.